A method for identifying abrasive grain signals based on a fused attention module and a parallel convolutional network

Abstract

This invention belongs to the field of oil wear particle monitoring technology, specifically relating to a wear particle signal identification method based on a fusion of attention modules and parallel convolutional networks. The method includes: establishing an adjustable set of feature parameters and a simulation data generation method using an accurate model of inductive oil wear particle detection signals; obtaining training data for a neural network model; and finally obtaining the wear particle signal identification result using the trained neural network. This invention's wear particle signal identification method based on a fusion of attention modules and parallel convolutional networks can achieve fast and accurate wear particle feature identification and classification by combining deep learning technology while avoiding distortion of target signal features. Furthermore, using convolutional neural networks for wear particle feature extraction avoids the difficulties in selecting threshold values ​​for various feature values ​​encountered with traditional feature extraction algorithms, reducing the requirements for users' expertise in signal processing.

Description

Technical Field

[0001] This invention belongs to the field of oil abrasive monitoring technology, specifically relating to an abrasive signal identification method based on a fused attention module and a parallel convolutional network. Background Technology

[0002] Online oil wear detection utilizes non-destructive sensors installed in lubricating oil pipelines to detect and analyze wear particles in oil in real time, thereby understanding the wear condition of key mechanical components and playing a crucial role in estimating the operating status and life prediction of complex equipment. With the continuous advancement of mechanical technology, many oil wear detection technologies based on different principles have been developed and are constantly being innovated. Inductive debris sensors (IDS) detect debris by sensing the non-periodic, regular voltage signal generated when particles pass through. They are popular among manufacturers and researchers in aero-engines, marine engines, and generator gearboxes due to their high sensitivity to ferromagnetic abrasive particles, ease of installation, low cost, and the fact that sensor performance is not easily affected by oil impurities. However, in practical applications, external electromagnetic interference and mechanical vibration inevitably cause the induced voltage signal acquired by IDS to contain a large amount of interference components, significantly impacting the accurate identification and extraction of wear particle signal characteristics and subsequent quantitative analysis.

[0003] To address this issue, numerous studies have explored signal processing techniques and feature extraction methods to improve the abrasive particle detection performance of IDS. Signal processing techniques primarily rely on modal decomposition, adaptive filtering, and bandwidth separation to enhance the signal-to-noise ratio of the detected signal, thereby suppressing the influence of non-abrasive signals on accurate abrasive particle identification. Feature extraction mainly includes binarization identification methods based on the geometric morphological features of the abrasive signal and a threshold in the transformation domain. The principle lies in the stable, noise-distinguishing local nonlinear characteristics of the abrasive-induced signal, often employing a segmentation-feature value calculation-threshold classification process for abrasive signal extraction. To assess the mechanical operating state, accurate counting of ferromagnetic abrasive particles of different sizes is essential for subsequent mechanical wear area analysis and wear degree evaluation. While these two main abrasive signal processing methods achieve good results in scenarios with sufficient prior knowledge, their over-reliance on signal-to-noise suppression leads to alterations in the local morphological features of the abrasive-induced voltage signal due to residual interference components or excessive filtering. Consequently, the preservation of abrasive features still falls short of the requirements for accurate abrasive feature extraction and analysis. Meanwhile, methods relying on simple nonlinear features can suffer from limitations in abrasive grain recognition due to factors such as the appropriateness of feature threshold selection and the high similarity between noise morphology features and abrasive grain features, posing challenges to the application and advancement of the technology. On the other hand, manual feature extraction and index discrimination are extremely complex processes, often requiring extensive professional knowledge and numerous field experiments. This results in the relatively low versatility of the aforementioned traditional methods, thus affecting the practicality and actual effectiveness of IDS. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes a method for identifying abrasive grain signals based on a fused attention module and a parallel convolutional network, comprising:

[0005] S1: The lubricating oil wear monitoring system, constructed using an inductive wear particle detection sensor, collects data on lubricating oil containing ferromagnetic wear particles to obtain the raw signal to be processed.

[0006] S2: Perform windowed discrete Fourier transform on the acquired raw signal to obtain the amplitude frequency domain spectrum of the entire signal. Then, calculate the amplitude, frequency and phase of each low-frequency harmonic interference by using iterative interpolation discrete Fourier transform algorithm and frequency domain compensation method to construct anti-phase components for harmonic interference suppression and generate simulation samples.

[0007] S3: Calculate the optimal number of stationary wavelet decomposition layers J based on the center frequency of the abrasive grains, divide the simulation samples into equal-length segments based on the optimal number of stationary wavelet decomposition layers J, extract the features of the abrasive grain signals and the simulation samples after equal-length segmentation to form a multi-scale feature training set, and complete the labeling process.

[0008] S4: The parallel convolutional network with fusion attention module is used to encode and reconstruct the J multi-scale feature vectors of a single sample in the training set, and the J reconstructed high-dimensional vectors are position-encoded and embedded. The dependency relationship between global and local information vectors is established by energy self-attention model to obtain the final classification vector, and Softmax is used to transform it into the predicted label probability.

[0009] S5: Construct a cross-entropy loss function based on the predicted label probability, fit the model using the cross-entropy loss function, and obtain the model with the best classification accuracy under multiple trainings as the abrasive signal identifier. Obtain a limited number of real abrasive signal samples with the proposed signal processing parameters, and enhance the actual detection performance of the identifier through feature transfer learning.

[0010] S6: Perform signal segmentation and multi-scale feature decomposition on the raw data acquired in real time, and input it into the post-transfer network model to obtain the abrasive signal identification results;

[0011] S7: Based on the accurate model of abrasive signal and key features of the sample, abrasive signal is reconstructed to obtain a noise-free abrasive detection signal, and finally abrasive feature extraction is achieved.

[0012] The beneficial effects of this invention are:

[0013] This invention achieves abrasive feature extraction by suppressing background noise in the acquired signal and then using parallel convolutional network operations, while maximizing the preservation of quantitative features of the abrasive signal. Currently, traditional algorithms based on signal decomposition and adaptive filtering struggle to balance abrasive feature protection with signal-to-noise ratio improvement. A prominent issue is that these algorithms alter the amplitude of the abrasive voltage signal, affecting the assessment of mechanical health and failing to provide reliable data for subsequent quantitative wear analysis. The abrasive signal identification method proposed in this invention, based on a fused attention module and parallel convolutional network, can achieve rapid and accurate abrasive feature identification and classification by combining deep learning technology while avoiding distortion of target signal features. Furthermore, using convolutional neural networks for abrasive feature extraction avoids the difficulties in selecting threshold values ​​for various features encountered with traditional feature extraction algorithms, reducing the requirements for users with expertise in signal processing. Attached Figure Description

[0014] Figure 1 This is a flowchart of a method for identifying abrasive signals based on a fused attention module and a parallel convolutional network according to the present invention.

[0015] Figure 2 This is a schematic diagram of the lubricating oil wear particle monitoring system of the present invention;

[0016] Figure 3 A schematic diagram showing the output of the raw signal to be processed by an inductive abrasive sensor.

[0017] Figure 4 This is a schematic diagram of the fusion attention module of the present invention;

[0018] Figure 5 This is a schematic diagram of the parallel convolutional network structure of the present invention;

[0019] Figure 6 This is a schematic diagram showing the original signal to be processed extracted by the abrasive monitoring system of the present invention and the signal after harmonic elimination;

[0020] Figure 7 This is a schematic diagram of the abrasive grain induced voltage signal identified by the parallel convolutional network of the present invention;

[0021] Figure 8 This is a schematic diagram of the reconstructed abrasive grain induced voltage signal of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] This invention provides a method for identifying abrasive grain signals based on a fused attention module and a parallel convolutional network, such as... Figure 1 As shown, it includes the following steps:

[0024] S1: The lubricating oil wear monitoring system, constructed using an inductive wear particle detection sensor, collects data on lubricating oil containing ferromagnetic wear particles to obtain the raw signal to be processed.

[0025] S2: Perform windowed discrete Fourier transform on the acquired raw signal to obtain the amplitude frequency domain spectrum of the entire signal. Then, calculate the amplitude, frequency and phase of each low-frequency harmonic interference by using iterative interpolation discrete Fourier transform algorithm and frequency domain compensation method to construct anti-phase components for harmonic interference suppression and generate simulation samples.

[0026] S3: Calculate the optimal number of stationary wavelet decomposition layers J based on the center frequency of the abrasive grains, divide the simulation samples into equal-length segments based on the optimal number of stationary wavelet decomposition layers J, extract the features of the abrasive grain signals and the simulation samples after equal-length segmentation to form a multi-scale feature training set, and complete the labeling process.

[0027] S4: The parallel convolutional network with fusion attention module is used to encode and reconstruct the J multi-scale feature vectors of a single sample in the training set, and the J reconstructed high-dimensional vectors are position-encoded and embedded. The dependency relationship between global and local information vectors is established by energy self-attention model to obtain the final classification vector, and Softmax is used to transform it into the predicted label probability.

[0028] S5: Construct a cross-entropy loss function based on the predicted label probability, fit the model using the cross-entropy loss function, and obtain the model with the best classification accuracy under multiple trainings as the abrasive signal identifier. Obtain a limited number of real abrasive signal samples with the proposed signal processing parameters, and enhance the actual detection performance of the identifier through feature transfer learning.

[0029] S6: Perform signal segmentation and multi-scale feature decomposition on the raw data acquired in real time, and input it into the post-transfer network model to obtain the abrasive signal identification results;

[0030] S7: Based on the accurate model of abrasive signal and key features of the sample, abrasive signal is reconstructed to obtain a noise-free abrasive detection signal, and finally abrasive feature extraction is achieved.

[0031] Specifically, the process of acquiring raw signals by the lubricating oil wear particle monitoring system includes:

[0032] Build as Figure 2 The diagram shows a platform schematic of a lubricating oil wear particle monitoring system, where the data acquisition card is model NI-9219 and the sampling frequency is set to f. s =5000, sampling time is 60s, and after removing subsequent redundant signals, abrasive particle detection signal with 280000 sampling points is obtained. The inductive abrasive particle detection sensor uses DC excitation, drive current I = 0.5A, preamplifier amplification factor is 2000, and peristaltic pump flow rate is set to 510ml / min. Abrasive particles (equivalent diameter of spheres) of 220μm, 73μm, 113μm, 65μm, 82μm, and three copper particles with an average diameter of 400μm are sequentially passed through the inductive abrasive particle detection sensor. The sensor will generate eight induced voltage signals similar to a single-cycle sine wave. The acquired raw signal to be processed is as follows: Figure 3 As shown, it can be seen that the raw signal to be processed collected in real time by the system not only contains the wear-particle induced voltage signal, but also includes harmonic interference caused by structural vibration and random noise caused by electromagnetic waves, etc. Therefore, it is impossible to quickly and accurately extract the morphological features of the tiny wear-particle induced voltage signal.

[0033] S2. Estimate the key features of the abrasive signal based on the sensor structure parameters and dynamic parameters of the deployment environment. Combine the precise model of the output signal of the inductive oil abrasive sensor to generate simulation data with adjustable feature parameters, and preprocess the original data and generated data.

[0034] Specifically, the acquired raw signal is first subjected to a windowed Discrete Fourier Transform (DFT) to obtain the amplitude frequency domain spectrum of the entire signal. Then, the amplitude, frequency, and phase of each low-frequency harmonic interference are calculated using an iterative interpolation DFT algorithm and a frequency domain compensation method to construct an anti-phase component for harmonic interference suppression. The generated data is simulated based on the estimated abrasive particle velocity, size, and sensor structural parameters. The simulation sample length N = 192, and the abrasive particle velocity is considered approximately equal to the oil velocity. Based on the frequency-velocity conversion formula, the abrasive particle frequency is approximately fp. d =106Hz. The specific process includes:

[0035] S21. Measure the length of the sensor's detection area and the oil velocity to approximate the abrasive grain center frequency. The calculation formula is:

[0036] f d =v d / l, where v d ≈v l

[0037] In the formula, f d v is the center frequency of the abrasive grains. d v is the abrasive particle velocity. l Let l be the oil flow velocity, and l be the length of the sensor detection area.

[0038] S22. Perform a windowed Fourier transform on the original signal y(n) to obtain the amplitude spectrum of the original signal;

[0039] S23. Extract the amplitude corresponding to all frequency points in the entire frequency band, perform discrete spectrum correction and normalization processing, extract spectral lines greater than 0.1 as harmonic signals, extract their amplitude and phase to construct an inverted continuous sinusoidal signal and superimpose it with the original signal to achieve harmonic interference suppression;

[0040] S24. Based on the estimated frequency f of the abrasive induced voltage signal d And estimate the amplitude {A1,A2,...,A} Q The model uses a wear-particle signal theoretical model to generate simulated wear-particle samples of different amplitudes in class Q. The wear-particle signal model is represented as:

[0041]

[0042] Where α is an adjustable parameter set according to the sensor coil structure parameters, Φ rThis is a simplified sensor structural parameter, where τ is the time delay at the center point of a single abrasive grain signal.

[0043] S25. Acquire sufficient non-abrasive signals using sensors, and perform preprocessing such as harmonic suppression with the same parameters to form noise samples; add Gaussian white noise to the generated data with varying signal-to-noise ratio to enhance the data, and finally complete the data preprocessing.

[0044] S3. Estimate the center frequency of the abrasive particles and calculate the optimal decomposition layer of the stationary wavelet. Perform equal-length simulation signal segmentation and extract the approximate coefficients of the J-layer decomposition of equal abrasive particles and noise samples to form a multi-scale feature training vector set and complete the labeling process.

[0045] Specifically, the steps for multi-scale feature extraction and training set construction include:

[0046] S31. Based on the principle of Discrete Wavelet Transform (DWT) decomposition, the number of sample stationary wavelet transform (SWT) decomposition levels is calculated using the following formula:

[0047] J = floor((lnf) s -lnαf d ) / ln2)-1

[0048] In the formula, f s J is the sampling frequency, and α is the expansion coefficient to prevent the abrasive characteristics from being destroyed during decomposition. Based on the abrasive frequency estimation, J is set to 4, and α is set to 2.

[0049] S32. For the u-th sample x u ∈R 1×N Where u = 1, 2, ..., U, the j-th layer features are represented as:

[0050] η u,j =Ψ a (x u ;j)

[0051] In the formula, Ψ a This represents the SWT approximation coefficient operator. Further, the original sample x... u The multi-scale feature vector can be represented as:

[0052] η u =[η u,1 ,η u,2 ,...η u,j ,...,η u,J ]

[0053] S33. By labeling the training samples, we can obtain the training set:

[0054] Y train ={η u |yu}

[0055] Among them, y u y is the label of the u-th sample. When the sample contains abrasive grain features, y = 0; when the sample is an interference signal sample, y = 1.

[0056] S4. The four multi-scale feature vectors of the u-th sample are encoded and reconstructed by a parallel convolutional network with a fusion attention module. The four reconstructed high-dimensional vectors are then encoded at their positions and the initial classification vector is inserted. The dependency relationship between the vectors is established by using a self-attention model. The final classification vector is obtained by combining global and local information.

[0057] like Figure 4 As shown, the process by which the attention module acquires global and local information is as follows:

[0058] S41: For the output feature map F = [f1, f2, ..., f...] of any layer in the neural network... i ,...,f I ], where f I For the i-th feature in the feature map, first calculate its channel global mean and global energy, and then obtain the fused channel attention. Its global mean and global energy are calculated using the following formula:

[0059]

[0060] Among them, z avg Let z be the global mean. pow Let M be the global energy, M be the total number of feature maps, and f be the global energy. i (m) is a vector f i The m-th instance in the array, where I is the total number of channels;

[0061] Global attention weights are obtained by fusing channel attention through vector concatenation and linear fusion:

[0062]

[0063] Among them, g c Here, δ represents the global attention weights, δ is the sigmoid activation function, Wc is the learnable linear transformation matrix, and C(·) represents the convolutional layer. For vector concatenation, k is the kernel size, β c This is the error term for the fully connected layer;

[0064] The global information vector is obtained by weighting the feature map channels based on global attention weights.

[0065] F'=F·g c +F

[0066] Where F' is the global information vector;

[0067] S42. Apply average pooling and energy pooling to the feature map from the vector dimension to obtain its local mean and local energy v, respectively. avg ∈R 1×M and v pow ∈R 1×M Furthermore, it uses a convolution operation with a kernel length of 7 to obtain its fused local attention weights: g s =δ(C s ([v avg ;v pow ],7)), where g s For local attention weights, C s (,7) represents a local convolution operation with a kernel length of 7;

[0068] The local information vector is obtained by weighting the feature map channels based on local attention weights:

[0069] F”=[f(i)·g s +f(i), i=1,2,...,I]

[0070] Where F” is the local information vector, f(i) is the i-th feature in the feature map, and · represents the Hadamard product.

[0071] like Figure 5 As shown, the parallel convolutional network combined with the fusion attention module consists of 5 layers of weight-shared stacked convolutional blocks, flattening layers, pooling layers, fully connected layers, and a one-dimensional Vision Transformer (ViT) self-attention classification layer. The weight-shared stacked convolutional blocks are one-dimensional convolution + layer normalization + ReLU activation + fusion attention module.

[0072] For input sample {η u ;y u First, multi-scale high-dimensional feature maps are extracted using stacked convolutional blocks with shared weights: H u =[h u,1 ,h u,2 ,...,h u,j ,...,h u,J ], where η u For sample x u Multi-scale feature vectors, y u Let H be the label of the u-th sample. u Indicates sample x u Multi-scale high-dimensional feature maps, h u,J For sample x u The J-th layer multi-scale high-dimensional features;

[0073] For H uEach instance is flattened separately, and mean pooling and max pooling with a kernel length of 4 and a stride of 4 are used to reduce the amount of data for each instance. Then, a shared weight fully connected layer is used to obtain the reconstructed multi-scale feature vector. in, For sample x u Reconstructing multi-scale feature vectors, For sample x u The J-th layer reconstructs multi-scale features;

[0074] The reconstructed multi-scale feature vectors are embedded with classification vectors and encoded at their positions. The results are then fed into the ViT model to model the global and local information dependencies of the multi-scale features. Finally, the classification is performed by a multilayer perceptron based on the fitted classification vectors.

[0075] Specifically, the final classification vector of any sample has a length of 2, and is transformed into predicted label probabilities using Softmax.

[0076] The structural parameters of the convolutional module in the parallel convolutional network are shown in Table 1:

[0077] Table 1. Structure parameters of the convolutional module in the parallel convolutional network

[0078]

[0079] The model was fitted using the cross-entropy loss function, and the model with the best classification accuracy under multiple training was used as the abrasive signal identifier. A limited number of real abrasive signal samples were obtained with the proposed signal processing parameters, and the actual detection performance of the identifier was enhanced through feature transfer learning.

[0080] Specifically, the cross-entropy function used in the example is defined as:

[0081]

[0082] Where, θ u It is the true label of the u-th sample. It is the predicted label of the u-th sample.

[0083] In this example, the real abrasive samples are first preprocessed using low-pass filtering and harmonic cancellation with the same preset parameters. Then, based on the estimated abrasive parameters, key parameters such as the width and amplitude of the feature signal are estimated to construct a sliding signal window. The sliding similarity of the real signals within this window is calculated to obtain the similarity feature domain. Based on the estimated environmental and structural parameters, 0.9 is selected as the similarity threshold, and indices greater than this threshold are extracted to locate the real abrasive signals. Finally, under the premise of complete abrasive features, real abrasive signal samples of length N=192 are extracted, and after labeling, multi-scale feature extraction is performed. The model is further fitted based on feature transfer learning.

[0084] S6. Perform signal segmentation and multi-scale feature decomposition on the raw data acquired in real time, and input it into the transfer network model to obtain the abrasive signal identification results;

[0085] Furthermore, the real-time data mentioned in step S6 is first preprocessed with the same parameters, and then a sliding segmentation is performed using a window of length N and stride N / 2 to obtain the samples to be identified. The sample prediction is completed by fitting a parallel convolutional network, and the non-abrasive signal segments are set to zero to finally achieve abrasive identification.

[0086] Specifically, the segmented signal length is N = 192, and the number of multi-scale eigendecomposition layers is J = 4. In this example, a detection signal with 250,000 sampling points and containing 8 abrasive grains is used for algorithm verification. The experimental signal is as follows: Figure 6 As shown, the identification results are as follows: Figure 7 As shown in the figure, a total of 8 abrasive grain induced voltage signals were accurately extracted, and the edge features of the abrasive grain signals were not destroyed.

[0087] S7. Based on the accurate model of abrasive signal and key features of the sample, abrasive signal is reconstructed to obtain noise-free abrasive detection signal, and finally abrasive feature extraction is achieved.

[0088] Furthermore, in step S7, the abrasive signal reconstruction first obtains the values ​​and indices of the maximum and minimum sample values, and then calculates the amplitude parameters. With width parameter Feature reconstruction is achieved based on a mathematical model of abrasive particle induced voltage. The specific parameter calculation process is as follows:

[0089]

[0090] The reconstructed abrasive grain signal, such as Figure 8 As shown, the amplitude characteristics of the abrasive signal are proportional to the size of the abrasive particles passing through the sensor sequentially, while the width characteristics remain stable when the flow rate is constant. Furthermore, by classifying and counting abrasive particles in different amplitude ranges, it is easier to analyze the abrasive particles individually later, thereby estimating the wear state of the mechanical equipment.

[0091] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for identifying abrasive grain signals based on a fused attention module and a parallel convolutional network, characterized in that, include: S1: The lubricating oil wear monitoring system, constructed using an inductive wear particle detection sensor, collects data on lubricating oil containing ferromagnetic wear particles to obtain the raw signal to be processed. S2: Perform windowed discrete Fourier transform on the acquired raw signal to obtain the amplitude frequency domain spectrum of the entire signal. Then, calculate the amplitude, frequency and phase of each low-frequency harmonic interference by using iterative interpolation discrete Fourier transform algorithm and frequency domain compensation method to construct anti-phase components for harmonic interference suppression and generate simulation samples. S3: Calculate the optimal number of stationary wavelet decomposition layers J based on the center frequency of the abrasive grains, divide the simulation samples into equal-length segments based on the optimal number of stationary wavelet decomposition layers J, extract the features of the abrasive grain signals and the simulation samples after equal-length segmentation to form a multi-scale feature training set, and complete the labeling process. S4: The parallel convolutional network with fusion attention module is used to encode and reconstruct the J multi-scale feature vectors of a single sample in the training set, and the J reconstructed high-dimensional vectors are position-encoded and embedded. The dependency relationship between global and local information vectors is established by energy self-attention model to obtain the final classification vector, and Softmax is used to transform it into the predicted label probability. The parallel convolutional network with fusion attention module consists of 5 layers of weight-shared stacked convolutional blocks, flattening layers, pooling layers, fully connected layers, and a one-dimensional Vision Transformer (ViT) self-attention classification layer. The weight-shared stacked convolutional blocks are one-dimensional convolution + layer normalization + ReLU activation + fusion attention module. S5: Construct a cross-entropy loss function based on the predicted label probability, fit the model using the cross-entropy loss function, and obtain the model with the best classification accuracy under multiple trainings as the abrasive signal identifier. Obtain a limited number of real abrasive signal samples with the proposed signal processing parameters, and enhance the actual detection performance of the identifier through feature transfer learning. S6: Perform signal segmentation and multi-scale feature decomposition on the raw data acquired in real time, and input it into the post-transfer network model to obtain the abrasive signal identification results; S7: Based on the accurate model of abrasive signal and key features of the sample, abrasive signal is reconstructed to obtain a noise-free abrasive detection signal, and finally abrasive feature extraction is achieved.

2. The abrasive particle signal identification method based on a fused attention module and a parallel convolutional network according to claim 1, characterized in that, The lubricating oil wear monitoring system constructed using an inductive wear particle detection sensor consists of a wear particle sensor, a signal acquisition card, an oil tank, an ultrasonic flow velocity sensor, a peristaltic pump, a DC power supply, a wear particle receiving device, and an experimental base. The signal acquisition card is a NI-9219 model, and the sampling frequency is set to... The sampling time is 60 seconds, and the number of sampling points obtained after removing subsequent redundant signals is: The abrasive particle detection signal is generated by the inductive abrasive particle detection sensor, which uses DC excitation with a drive current of I = 0.5A, a preamplifier with an amplification factor of 2000, and a peristaltic pump flow rate of 510ml / min. The abrasive particles with equivalent spherical diameters of 220μm, 73μm, 113μm, 65μm, 82μm, and three copper particles with an average diameter of 400μm pass through the inductive oil abrasive particle detection sensor in sequence, generating eight induced voltage signals similar to a single-cycle sine wave.

3. The abrasive particle signal identification method based on a fused attention module and a parallel convolutional network according to claim 1, characterized in that, The acquired raw signal is subjected to a windowed Discrete Fourier Transform (DFT) to obtain the amplitude frequency domain spectrum of the entire signal. Then, the amplitude, frequency, and phase of each low-frequency harmonic interference are calculated using an iterative interpolation DFT algorithm and a frequency domain compensation method to construct an inverse component for harmonic interference suppression. Simulation samples are generated, including: S21: Measure the length of the sensor's detection area and the oil velocity to approximate the abrasive grain center frequency. ,in, The abrasive grain center frequency, For abrasive particle velocity, , The oil flow rate, The length of the sensor's detection area; S22: Raw signal to be processed The original signal amplitude spectrum is obtained by performing a windowed Fourier transform; S23: Extract the amplitude corresponding to all frequency points in the entire frequency band, perform discrete spectrum correction and normalization processing, extract spectral lines greater than 0.1 as harmonic signals, extract their amplitude and phase to construct an inverted continuous sine signal and superimpose it with the original signal to achieve harmonic interference suppression and generate simulated abrasive particle samples. S24: Based on the estimated frequency of the abrasive induced voltage signal and estimated amplitude The model uses a wear particle signal theoretical model to generate simulated wear particle samples of different amplitudes in class Q. S25: Acquire sufficient non-abrasive signals using sensors, and generate simulated non-abrasive samples by performing harmonic suppression with the same parameters; add Gaussian white noise to the generated data with varying signal-to-noise ratio for data enhancement, and finally complete the data preprocessing.

4. The abrasive particle signal identification method based on a fused attention module and a parallel convolutional network according to claim 3, characterized in that, The theoretical model of abrasive signals includes: in, For the theoretical model of abrasive signals, These are adjustable parameters set according to the sensor coil structure parameters. To simplify the sensor's structural parameters, The delay is the signal center point delay for a single abrasive grain, where n is the number of signal points. Let be the amplitude of the q-th abrasive grain induced voltage signal. The sampling frequency.

5. The abrasive grain signal identification method based on a fused attention module and a parallel convolutional network according to claim 1, characterized in that, The optimal number of stationary wavelet decomposition layers J is calculated based on the abrasive grain center frequency, including: in, Sampling frequency, To prevent the abrasive grain characteristics from being destroyed during decomposition, the expansion coefficient is set to 4 based on the abrasive grain estimated frequency wavelet maximum decomposition level J and the frequency band expansion coefficient. Take 2, The abrasive grain center frequency is denoted as .

6. The abrasive particle signal identification method based on a fused attention module and a parallel convolutional network according to claim 1, characterized in that, The features of the extracted abrasive grain signals and the simulated samples after equal-length segmentation are used to construct a multi-scale feature training set, which is then labeled, including: A sample set is formed by combining equal amounts of abrasive particle signals with simulated sample fragments. For the u-th sample in the sample set... Where u = 1, 2, ..., U, the j-th layer features are represented as: Its multi-scale eigenvectors are represented as: Where U is the sample size. For the sample The j-th layer features, This is the SWT approximation coefficient operator. For the sample Multi-scale feature vectors For the sample The J-th layer features; Perform sample labeling: ,in, For labeled samples, Let be the label of the u-th sample. When the sample contains abrasive grain features, When the sample is an interference signal sample, Label all samples in the sample set to obtain the final training set.

7. The abrasive grain signal identification method based on a fused attention module and a parallel convolutional network according to claim 1, characterized in that, A parallel convolutional network incorporating a fusion attention module is used to encode and reconstruct J multi-scale feature vectors from a single sample in the training set. The output J reconstructed high-dimensional vectors are then positionally encoded and embedded. An energy self-attention model is used to establish the dependency relationship between global and local information vectors, resulting in the final classification vector, including: For input samples First, multi-scale high-dimensional feature maps are extracted using stacked convolutional blocks with shared weights: ,in, For the sample Multi-scale feature vectors Let be the label of the u-th sample. Indicates sample Multi-scale high-dimensional feature maps. For the sample The J-th layer multi-scale high-dimensional features; right Each instance is flattened separately, and mean pooling and max pooling with a kernel length of 4 and a stride of 4 are used to reduce the amount of data for each instance. Then, a shared weight fully connected layer is used to obtain the reconstructed multi-scale feature vector. ,in, For the sample Reconstructing multi-scale feature vectors, For the sample The J-th layer reconstructs multi-scale features; The reconstructed multi-scale feature vectors are embedded with classification vectors and encoded at their positions. The results are then fed into the ViT model to model the global and local information dependencies of the multi-scale features. Finally, the classification is performed by a multilayer perceptron based on the fitted classification vectors.

8. The abrasive grain signal identification method based on a fused attention module and a parallel convolutional network according to claim 7, characterized in that, The reconstructed multi-scale feature vectors are subjected to classification vector embedding and positional encoding, and then fed into the ViT model to model the global and local information dependencies of multi-scale features, including: in, Output feature maps for the ViT module. For Transformer layer operators in the ViT model, For the sample Reconstructing multi-scale feature vectors, To insert the classification encoding vector, This is a classification vector.

9. The abrasive particle signal identification method based on a fused attention module and a parallel convolutional network according to claim 1, characterized in that, The global and local information vectors include: S41: Output feature maps for any layer of the neural network ,in, For the i-th feature in the feature map, first calculate its channel global mean and global energy, and then obtain the fused channel attention. Its global mean and global energy are calculated using the following formula: in, The global mean. For global energy, The total number of feature maps, For vector f i The m-th instance in Total number of channels; Global attention weights are obtained by fusing channel attention through vector concatenation and linear fusion: in, For global attention weights, It is the Sigmoid activation function. For learnable linear transformation matrices, It is a convolutional layer. For vector concatenation, The kernel size is [size]. This is the error term for the fully connected layer; The global information vector is obtained by weighting the feature map channels based on global attention weights. in, Global information vector; S42. Apply average pooling and energy pooling to the feature map from the vector dimension to obtain its local mean and local energy, respectively. and Furthermore, it uses a convolution operation with a kernel length of 7 to obtain its fused local attention weights: ,in, For local attention weights, For local convolution operations with a kernel length of 7; The local information vector is obtained by weighting the feature map channels based on local attention weights: in, For local information vectors, For the i-th feature in the feature map, It represents the Hadamardi (or Hadama) stack.

10. The abrasive particle signal identification method based on a fused attention module and a parallel convolutional network according to claim 1, characterized in that, The cross-entropy loss function includes: in, Let U be the cross-entropy loss function, and U be the number of samples. Let u be the true label of the u-th sample. Let be the predicted label for the u-th sample.

Images